AI Process Log
How I actually work with AI.
An honest, ongoing log of where AI accelerates the work, where human judgment still does the deciding, and what I'm learning as I go.
Most AI portfolios show outputs. This one shows process. Each entry is a real moment from my work, including what I prompted, what worked, what didn't, and what I'd do differently next time. It's a public version of my private notes, because I think the most valuable part of working with AI right now is watching how other people are actually doing it.
June 24, 2026
Debugging
Diagnosing a parent route bug and locking the Frameworks architecture
Restructured the site information architecture so Frameworks is the parent category and Tools are nested under each framework. Created three placeholder framework pages for in-development frameworks - three placeholder pages with the right information architecture - and verified the parent route fix by navigating to each framework page and confirming it rendered its own content instead of falling back to the Frameworks index. Added a Frameworks dropdown to the top navigation so visitors can browse frameworks from any page.
- Input
- A broken routing pattern where clicking into specific framework cards kept returning visitors to the Frameworks index page. Multiple Lovable patch attempts had already failed to fix it.
- Output
- A restructured site information architecture with Frameworks as the parent category and Tools nested under each framework. Three new placeholder pages for in-development frameworks (Operating Cadence, Meeting Design, Change Cadence). A working Frameworks dropdown in the top navigation. A fully fixed parent route in TanStack Router so child framework pages now render correctly.
- Safeguards
- Refused to accept Lovable's confident 'this is fixed' claims when the bug clearly persisted. Forced a stop on the patching cycle and required Lovable to produce a structured diagnostic report instead. The diagnostic surfaced the actual root cause (missing Outlet in parent route), which turned a one-hour debugging spiral into a single precise fix.
What I learned
The fastest way to stop a debugging spiral is to stop accepting the tool's diagnosis and demand a structured diagnostic of your own. This applies to any AI tool, not just Lovable.
June 23, 2026
System Design
Restructuring the site information architecture and completing the AI Adoption Readiness framework
Today's work split into three: overhauling the site information architecture into a clearer Frameworks and Tools structure, adding the Operating Cadence Diagnostic tool with a full input form and AI-generated output, and completing the first iteration of all six dimensions of the AI Adoption Readiness framework.
- Input
- A working brief covering the site restructure, Operating Cadence Diagnostic tool requirements, and the final dimension content for the AI Adoption Readiness framework.
- Output
- A restructured site with clear Frameworks and Tools sections, a live Operating Cadence Diagnostic with full form and AI output, and all six dimensions of the AI Adoption Readiness framework published.
- Safeguards
- Manually verified every navigation link and cross-reference after restructuring, confirmed the site index captured all new pages, and checked timestamps and status badges across every affected page.
What I learned
Connective elements drift quietly. The site index sat outdated for two sessions before we noticed. Individual updates landed cleanly, but the elements that touch everything (site index, navigation references, status badges, timestamps) fell behind. The pattern: when you change something specific, also check what references it indirectly. Project instructions now include a standing rule to flag updates to the site index whenever a new framework, tool, or page is added.
May 25, 2026
System Design
Building a portfolio site with Claude and Lovable as a connected workflow
Spent the day designing and building a full portfolio system: a GitHub repo for canonical content, a Lovable site for the public-facing experience, and a process log to capture the thinking behind both. The goal wasn't to ship a site. It was to demonstrate how I think about systems.
Claude held the strategy and scaffolding. Lovable executed the build. The combination let me move much faster than either tool alone would have allowed, while still keeping me in control of the decisions.
- Input
- Site goals, audience notes, content inventory, and a sketch of the information architecture pasted into Claude as a single brief.
- Output
- A repo structure, route plan, and copy scaffolding that Lovable then turned into a working site I edited section by section.
- Safeguards
- I reviewed every generated section before it shipped, kept canonical content in GitHub so nothing lived only inside a tool, and refused any copy I couldn't say in my own voice.
What I learned
Tool friction is real and worth noticing. The GitHub connection bug with Lovable cost me an hour, but it also generated content for this log. The friction itself became material.
May 25, 2026
Sensemaking
Designing the AI Adoption Readiness framework
Used Claude to research the current state of AI adoption (sources: Stanford AI Index, BCG, Gartner, HCLTech). The data was bleak: 48% of orgs call AI adoption a disappointment, 60% of projects expected to be abandoned by end of 2026.
The temptation was to start prescribing solutions. Instead, I worked backward from the failure patterns to identify what dimensions actually determine readiness. Landed on six. The framework that came out of it is now live on this site.
- Input
- A research prompt naming the question, the sources I trusted (Stanford AI Index, BCG, Gartner, HCLTech), and an instruction to extract failure patterns rather than recommendations.
- Output
- A clustered summary of why AI initiatives stall, which I shaped into the six readiness dimensions now on the site.
- Safeguards
- Verified every statistic against the original source, threw out anything I couldn't trace, and wrote the framework myself rather than letting Claude propose the dimensions.
What I learned
Frameworks should be designed backward from observed failure, not forward from best-practice theory. The dimensions that matter are the ones that explain why things break, not the ones that sound smart.
May 17, 2026
Synthesis
Turning 40 interview transcripts into a single map
Used Claude to cluster themes from raw transcripts, then hand-edited the map. AI did the heavy reading; the judgment stayed with me.
- Input
- 40 anonymized transcripts plus a coding rubric describing the themes I cared about and the ones I explicitly did not.
- Output
- A first-pass thematic map with quote-level citations back to each transcript, which I edited down to the version we used with the team.
- Safeguards
- Stripped names and identifying details before anything touched the model, kept transcripts out of any tool with training on by default, and spot-checked every cluster against the underlying quotes.
What I learned
AI is great at pattern recognition across volume. It's bad at deciding what matters. The division of labor that worked: AI surfaced everything, I decided what to keep.
May 10, 2026
Drafting
First-pass operating model documentation
Prompted Claude with structure and constraints, refined output through three review cycles with the leadership team. Cut drafting time roughly in half.
- Input
- An outline of the operating model, the constraints it had to respect, and examples of the tone and length I wanted.
- Output
- A first draft of the documentation I edited with the leadership team across three review cycles before it went internal.
- Safeguards
- Used only non-confidential framing in the prompt, had a human owner sign off on each section, and flagged anything that read as too generic to be useful.
What I learned
First drafts are where AI saves the most time. Final drafts still need a human who knows the politics of the room.
New entries added regularly. This is a working document, not a polished portfolio.
Last updated: June 24, 2026